To add a bit to Parth's answer it has been trained using metric learning type losses and normal softmax combined. There is a next version of combined low/high level features combined algorithms we are slowly going to update into.